The shift that mattered most was not better code generation. It was moving from reusable commands and templates into a workflow that can carry context, trigger the right behavior, and verify work automatically.

A recent feature made the difference obvious. I used Codex to build interactive skill selection for "ai-devkit skill add". I gave one sentence of instruction, and the workflow carried the task through requirements, design, planning, implementation, verification, tests, and code review.

The whole session took under an hour. The actual feature flow was around 30 minutes.

What I found interesting was not just that AI wrote code. It was that the workflow left behind requirements, design docs, planning artifacts, tests derived from requirements, and verification against the spec instead of just a diff.

A few things that felt important in practice:

- memory pulled back an old CLI rule I had forgotten I stored - review phases could loop backward instead of blindly moving forward - verification caught drift between implementation and design - I still made the product decisions and fixed the last failing test myself

I am curious how others here are thinking about this.

Are you mostly optimizing prompts, or are you now trying to optimize the workflow layer around the model?